IIGROUP Submissions for WMT22 Word-Level AutoCompletion Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1187 - 1191 December 7-8, 2022 ©2022 Association for Computational Linguistics IIGROUP Submissions for WMT22 Word-Level AutoCompletion Task Cheng Yang Siheng Li Chufan Shi Yujiu Yang Tsinghua Shenzhen International Graduate School, Tsinghua University {yangc21,lisiheng21,scf22}@mails.tsinghua.edu.cn yang.yujiu@sz.tsinghua.edu.cn Abstract This paper presents IIGroup’s submission to the WMT22 Word-Level AutoCompletion(WLAC) Shared Task in four language directions. We propose to use a Generate-then-Rerank frame- work to solve this task. More specifically, the generator is used to generate candidate words and recall as many positive candidates as pos- sible. To facilitate the training process of the generator, we propose a span-level mask predic- tion task. Once we get the candidate words, we take the top-K candidates and feed them into the reranker. The reranker is used to select the most confident candidate. The experimental results in four language directions demonstrate the effectiveness of our systems. Our systems achieve competitive performance ranking 1st in English to Chinese subtask and 2nd in Chinese to English subtask. 1 Introduction Recent advances in neural machine translation (Bahdanau et al., 2015; Vaswani et al., 2017) al- low us to generate high-quality translation results. However, as it’s pointed out by Li et al. (2021) that in some scenarios(e.g., legal instruments), the re- sults of machine translation can’t directly replace human translations due to their imperfections(e.g., terminology translation error). Therefore, more and more researchers pay attention to Computer- aided translation(CAT)(Barrachina et al., 2009; Santy et al., 2019; Huang et al., 2021; Xiao et al., 2022), which focuses on leveraging the advantages of NMT systems to increase the effectiveness and efficiency of the human translation process. To further promote the development of CAT, WMT22 proposes a novel task —— Word-Level AutoCompletion(WLAC). In the Word-Level Auto- Completion task, given a source sentence x, target context and human-typed characters t, an ideal sys- tem is expected to be able to predict the target word w that should be placed in the target context. We participate in the WMT22 shared Word- Level AutoCompletion task in four language di- rections: Chinese ⇒ English, English ⇒ Chinese, German ⇒ English and English ⇒ German and submit a system for each language direction. We develop a Generate-then-Rerank framework- based system for each language direction. Based on the vanilla Transformer architecture, we adopt a bidirectional decoder, which can predict the cur- rent target word by paying attention to the source sentence and both the left-side and right-side target context. The paper is organized as follows, section 2 gives the overview of the data used in the shared task and preprocessing operations for the data, while sec- tion 3 describes our training techniques, including model architecture, span-level mask prediction, etc. Section 4 presents our experimental results. Finally, our conclusions are summarized in Section 5. 2 Data In this section, we first introduce the datasets used to train our systems, then we introduce how to prepare the simulated training data for the WLAC shard task and describe the vocabulary for each language direction. …

Content truncated to 500 words